A deep neural network that learns from large-scale experimental optical data—rather than repeated electromagnetic simulations—can rapidly predict device performance, explore large design spaces, and perform inverse design in real time (without repeatedly running computationally expensive simulations).